Merge branch 'main' into add-vgf-ops-tests-part3

YufengShi-dudu · web-flow · commit 821d0f29cb4e · 2025-08-06T15:12:14.000+01:00
diff --git a/backends/nxp/backend/ir/converter/node_converters/ops_converters/sigmoid_converter.py b/backends/nxp/backend/ir/converter/node_converters/ops_converters/sigmoid_converter.py
@@ -1,5 +1,4 @@
 # Copyright 2025 NXP
-# All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
diff --git a/backends/nxp/backend/ir/converter/node_converters/shared/conv_utils.py b/backends/nxp/backend/ir/converter/node_converters/shared/conv_utils.py
@@ -1,8 +1,7 @@
 # Copyright 2023-2025 NXP
 #
-# License: LA_OPT_NXP_Software_License
-# See the LICENSE_LA_OPT_NXP_Software_License for more details.
-#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
 
 from copy import copy
 from dataclasses import dataclass
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/add_n_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/add_n_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/add_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/add_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/average_pool_2d_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/average_pool_2d_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/leaky_relu_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/leaky_relu_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/log_softmax_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/log_softmax_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/max_pool_2d_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/max_pool_2d_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/reshape_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/reshape_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/softmax_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/softmax_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/sub_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/sub_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/backend/ir/tflite_generator/builtin_options/transpose_options.py b/backends/nxp/backend/ir/tflite_generator/builtin_options/transpose_options.py
@@ -1,5 +1,6 @@
 #
 # Copyright 2023 Martin Pavella
+# Copyright 2024 NXP
 #
 # License: MIT
 # See the LICENSE_MIT for more details.
diff --git a/backends/nxp/neutron_partitioner.py b/backends/nxp/neutron_partitioner.py
@@ -1,4 +1,4 @@
-# Copyright (c) 2024-2025 NXP
+# Copyright 2024-2025 NXP
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
diff --git a/backends/nxp/quantizer/patterns.py b/backends/nxp/quantizer/patterns.py
@@ -279,7 +279,7 @@ def partition_types(self):
         return [torch.ops.aten.flatten.using_ints]
 
 
-class HardTanhPattern(SharedSpecPattern):
+class HardTanhPattern(QuantizationPattern):
     """
     Quantizer for HardTanh operator. Shared quantization spec is selected, as activation functions usually follows
     computation layer.
@@ -288,8 +288,23 @@ class HardTanhPattern(SharedSpecPattern):
     def partition_types(self):
         return [torch.ops.aten.hardtanh.default]
 
+    def get_anchors(
+        self, gm: fx.GraphModule, fused_partition: List[fx.GraphModule]
+    ) -> PartitionAnchors | None:
+        node = fused_partition[0].nodes[-1]
+
+        return PartitionAnchors(
+            inputs=[(node, 0)],
+            weights=[],
+            biases=[],
+            output=[(node,)],
+        )
+
+    def replacement_op(self):
+        raise AssertionError()
+
 
-class HardTanhInPlacePattern(SharedSpecPattern):
+class HardTanhInPlacePattern(QuantizationPattern):
     """
     Quantizer for HardTanh operator with param inplace=True. Shared quantization spec is selected, as activation
     functions usually follows computation layer.
@@ -298,6 +313,21 @@ class HardTanhInPlacePattern(SharedSpecPattern):
     def partition_types(self):
         return [torch.ops.aten.hardtanh_.default]
 
+    def get_anchors(
+        self, gm: fx.GraphModule, fused_partition: List[fx.GraphModule]
+    ) -> PartitionAnchors | None:
+        node = fused_partition[0].nodes[-1]
+
+        return PartitionAnchors(
+            inputs=[(node, 0)],
+            weights=[],
+            biases=[],
+            output=[(node,)],
+        )
+
+    def replacement_op(self):
+        raise AssertionError()
+
 
 class LinearPattern(QuantizationPattern):
     def partition_types(self) -> List[OpOverload]:
diff --git a/backends/nxp/tests/executorch_pipeline.py b/backends/nxp/tests/executorch_pipeline.py
@@ -34,6 +34,15 @@ def _quantize_model(model, calibration_inputs: list[tuple[torch.Tensor]]):
     return m
 
 
+def get_random_float_data(input_shapes: tuple[int] | list[tuple[int]]):
+    # TODO: Replace with something more robust.
+    return (
+        (torch.randn(input_shapes),)
+        if type(input_shapes) is tuple
+        else tuple(torch.randn(input_shape) for input_shape in input_shapes)
+    )
+
+
 def to_quantized_edge_program(
     model: torch.nn.Module,
     input_shapes: tuple[int] | list[tuple[int]],
@@ -47,12 +56,7 @@ def to_quantized_edge_program(
             "For multiple inputs, provide" " list[tuple[int]]."
         )
 
-    random_tensors = (
-        (torch.randn(input_shapes),)
-        if type(input_shapes) is tuple
-        else tuple(torch.randn(input_shape) for input_shape in input_shapes)
-    )
-    calibration_inputs = [random_tensors, random_tensors]
+    calibration_inputs = [get_random_float_data(input_shapes) for _ in range(4)]
     example_input = (
         (torch.ones(input_shapes),)
         if type(input_shapes) is tuple
diff --git a/backends/nxp/tests/ir/converter/node_converter/test_hardtanh_converter.py b/backends/nxp/tests/ir/converter/node_converter/test_hardtanh_converter.py
@@ -39,7 +39,7 @@ def forward(self, x):
         return self.block(x)
 
 
-class CustomHardTanhBlock(torch.nn.Module):
+class ConvHardTanhBlock(torch.nn.Module):
     def __init__(
         self,
         conv_in_channels: int = 3,
@@ -89,16 +89,18 @@ def test_relu6_quant(mocker, input_shape: tuple[int], inplace: bool):
     )
 
 
-@pytest.mark.parametrize("input_shape", [(1, 3, 128, 128), (1, 3, 256, 256)])
+@pytest.mark.parametrize("input_shape", [(1, 3, 16, 16), (1, 3, 32, 32)])
 @pytest.mark.parametrize(
     "activation_range", list(HardTanhConverter.supported_modes_map.keys())
 )
 @pytest.mark.parametrize("inplace", [True, False])
 def test_custom_hardtanh_quant(
     mocker, input_shape: tuple[int], activation_range: tuple[int, int], inplace: bool
 ):
+    # TODO(13063): This test suffers from non-ideal testing random quantization, because we always use range <0,1>.
+    #  We should update (decrease atol) when the Conv/Linear + Activation fuse at quantization is in place.
     min_val, max_val = activation_range
-    model = CustomHardTanhBlock(
+    model = ConvHardTanhBlock(
         conv_in_channels=input_shape[1],
         min_act_val=min_val,
         max_act_val=max_val,
@@ -122,5 +124,5 @@ def test_custom_hardtanh_quant(
         tflite_input_preprocess=ToNHWCPreprocess(),
         tflite_output_preprocess=ToNCHWPreprocess(),
         input_data=input_data,
-        atol=1.0,
+        atol=2.0,
     )
diff --git a/backends/nxp/tests/ir/converter/node_converter/test_sigmoid_converter.py b/backends/nxp/tests/ir/converter/node_converter/test_sigmoid_converter.py
@@ -1,5 +1,4 @@
 # Copyright 2025 NXP
-# All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
diff --git a/backends/nxp/tests/ir/edge_passes/test_remove_io_quant_ops_pass.py b/backends/nxp/tests/ir/edge_passes/test_remove_io_quant_ops_pass.py
@@ -1,7 +1,8 @@
-# Copyright 2025 NXP
+# Copyright 2024-2025 NXP
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
+
 import itertools
 
 import executorch.kernels.quantized  # noqa F401
diff --git a/backends/transforms/remove_clone_ops.py b/backends/transforms/remove_clone_ops.py
@@ -6,26 +6,45 @@
 
 # pyre-strict
 
+from typing import Set
+
 import torch
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
+from executorch.exir.passes import dead_code_elimination_pass
+from executorch.exir.passes.remove_noop_pass import _DEQUANT_OPS, eliminate_dq_q
 
 
-def remove_clone_ops(graph: torch.fx.Graph) -> torch.fx.Graph:
+class RemoveCloneOpsTransform(ExportPass):
     """
-    Remove clone op nodes and replace uses with parent node.
+    Trim the 'identity' operators to reduce the unnecessary copy overhead.
     """
-    clone_op = exir_ops.edge.aten.clone.default
-    for node in graph.nodes:
-        if node.op == "call_function" and node.target == clone_op:
-            with graph.inserting_after(node):
-                node.replace_all_uses_with(node.args[0])
 
-    graph.eliminate_dead_code()
-    return graph
+    clone_ops: Set[torch._ops.OpOverload] = {
+        exir_ops.edge.aten.clone.default,
+    }
 
+    def __init__(self) -> None:
+        super().__init__()
+
+    def _remove(self, graph_module: torch.fx.GraphModule) -> None:
+        dequant_nodes = []
+
+        for n in graph_module.graph.nodes:
+            if n.target not in self.clone_ops:
+                continue
+
+            to_be_remove = n
+            for user_n in list(n.users.keys()):
+                user_n.replace_input_with(n, n.args[0])
+            if n.args[0].target in _DEQUANT_OPS:
+                dequant_nodes += [n.args[0]]
+            graph_module.graph.erase_node(to_be_remove)
+
+        eliminate_dq_q(graph_module, dequant_nodes)
 
-class RemoveCloneOpsTransform(ExportPass):
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
-        graph_module.graph = remove_clone_ops(graph_module.graph)
+        self._remove(graph_module)
+        graph_module.recompile()
+        dead_code_elimination_pass(graph_module)
         return PassResult(graph_module, True)
diff --git a/backends/transforms/targets.bzl b/backends/transforms/targets.bzl
@@ -109,6 +109,7 @@ def define_common_targets():
         srcs = ["remove_clone_ops.py"],
         visibility = [
             "//executorch/backends/...",
+            "@EXECUTORCH_CLIENTS",
         ],
         deps = [
             "//caffe2:torch",
@@ -242,3 +243,15 @@ def define_common_targets():
             ":rank_0_to_rank_1",
         ],
     )
+
+    runtime.python_test(
+        name = "test_remove_clone_ops",
+        srcs = [
+            "test/test_remove_clone_ops.py",
+        ],
+        deps = [
+            "//caffe2:torch",
+            "//executorch/exir:lib",
+            ":remove_clone_ops",
+        ],
+    )
diff --git a/backends/transforms/test/test_remove_clone_ops.py b/backends/transforms/test/test_remove_clone_ops.py
diff --git a/runtime/executor/method.cpp b/runtime/executor/method.cpp

Original file line number	Diff line number	Diff line change
`@@ -1,5 +1,4 @@`
`1`	`1`	`# Copyright 2025 NXP`
`2`		`-# All rights reserved.`
`3`	`2`	`#`
`4`	`3`	`# This source code is licensed under the BSD-style license found in the`
`5`	`4`	`# LICENSE file in the root directory of this source tree.`
Original file line number	Diff line number	Diff line change
`@@ -1,5 +1,6 @@`
`1`	`1`	`#`
`2`	`2`	`# Copyright 2023 Martin Pavella`
	`3`	`+# Copyright 2024 NXP`
`3`	`4`	`#`
`4`	`5`	`# License: MIT`
`5`	`6`	`# See the LICENSE_MIT for more details.`